Semi-supervised Quality Evaluation of Colonoscopy Procedures

TL;DR

This paper tackles the polyp miss rate in colonoscopy by introducing unsupervised, video-driven quality metrics that quantify per-procedure inspection quality. It builds an end-to-end pipeline based on contrastive self-supervised learning to obtain frame embeddings, clusters frames into informative groups, and derives both online (real-time) and offline (post-procedure) quality metrics tied to polyp detection outcomes. Through a Bayesian analysis, the authors connect the online quality metric to polyp-detection probability $P(D|E,Q)$ and demonstrate strong alignment with polyp detection sensitivity, while the offline metric $Q_{\text{Offline}}$ correlates with the standard PPC metric. The findings suggest real-time quality feedback during withdrawal and reliable post-procedure quality assessment, with potential to improve ADR in future prospective studies.

Abstract

Colonoscopy is the standard of care technique for detecting and removing polyps for the prevention of colorectal cancer. Nevertheless, gastroenterologists (GI) routinely miss approximately 25% of polyps during colonoscopies. These misses are highly operator dependent, influenced by the physician skills, experience, vigilance, and fatigue. Standard quality metrics, such as Withdrawal Time or Cecal Intubation Rate, have been shown to be well correlated with Adenoma Detection Rate (ADR). However, those metrics are limited in their ability to assess the quality of a specific procedure, and they do not address quality aspects related to the style or technique of the examination. In this work we design novel online and offline quality metrics, based on visual appearance quality criteria learned by an ML model in an unsupervised way. Furthermore, we evaluate the likelihood of detecting an existing polyp as a function of quality and use it to demonstrate high correlation of the proposed metric to polyp detection sensitivity. The proposed online quality metric can be used to provide real time quality feedback to the performing GI. By integrating the local metric over the withdrawal phase, we build a global, offline quality metric, which is shown to be highly correlated to the standard Polyp Per Colonoscopy (PPC) quality metric.

Figures (5)

• Figure 1: Method overview.(Left) Two augmented views for each frame are used to train the encoder and the projection head using contrastive learning. (Right top) Feature representations are directly clustered into semantically meaningful groups using K-means. (Right middle) Learning clusters' associations. (Right bottom) At inference time, cluster attributes are leveraged for quality metric evaluation.
• Figure 2: T-SNE plot of frame embeddings.$K$-means clusters are color coded.
• Figure 3: Clusters visualization. Random selection of frames from each cluster.
• Figure 4: The likelihood of detecting an existing polyp in a short video segment as a function of local quality metric $Q$.
• Figure 5: $\mathbf{Q}_{\text{Offline}}$ during the withdrawal phase. (Left) The relationship between the proposed offline quality measure and the actual number of polyps detected, when $Q_{\text{Offline}}$ observations are divided into five equal-sized groups. (Right) Procedures with high $Q_{\text{Offline}}$ values are likely to have polyps.